Testifier Audit Report

Baseline Profile

Chart Help

What is this? Baseline trend of submissions and pro share. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Volume, Pro rate, Wilson low / Wilson high, Low-power.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Baseline trend of submissions and pro share.

Volume: Bars show record volume in each time bucket.
Pro rate: Line shows pro-position share per bucket.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Baseline Profile focuses on baseline volume and composition drift. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Baseline breaks in volume or composition that align with detector flags and repeat across adjacent windows. Short, isolated spikes in volume with no matching shift in pro rate or corroborating detector flags are often random campaign pulses rather than systemic manipulation. Extended level shifts (for example, 60-240 minutes) in both volume and composition, especially when Wilson bands tighten, suggest a meaningful regime change worth cross-checking against changepoints and composite evidence. Very low overnight volume can create dramatic percentage swings; prioritize windows where elevated rates persist after local volume recovers into daytime traffic. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate a short notice event, reminder blast, or temporary queue release. Momentary lows can indicate normal minute-level quiet periods or ingest timing jitter. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate a sustained participation regime shift that can affect all downstream detectors. Extended lows can indicate potential ingestion gaps, hearing lulls, or sustained reduced campaign activity. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Short, isolated spikes in volume with no matching shift in pro rate or corroborating detector flags are often random campaign pulses rather than systemic manipulation.
Extended level shifts (for example, 60-240 minutes) in both volume and composition, especially when Wilson bands tighten, suggest a meaningful regime change worth cross-checking against changepoints and composite evidence.
Very low overnight volume can create dramatic percentage swings; prioritize windows where elevated rates persist after local volume recovers into daytime traffic.

Common benign causes: Hearing schedule transitions and reminder cascades can create expected baseline shifts.

`baseline_day_hour_volume`

Chart Help

What is this? Day/hour baseline heatmap. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Cell color, X/Y axes.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Day/hour baseline heatmap.

Cell color: Darker cells indicate higher submission volume for that weekday/hour.
X/Y axes: X-axis is hour of day; Y-axis is weekday.

`baseline_top_names`

Chart Help

What is this? Top-frequency names. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis names.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Top-frequency names.

Bar height: Total submissions associated with each displayed name.
X-axis names: Most frequent names (trimmed to top slice for readability).

`baseline_name_length_distribution`

Chart Help

What is this? Name-length histogram view. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Name-length histogram view.

Bar height: Count of names with the corresponding character length.
X-axis: Normalized name length in characters.

Burst Windows

Chart Help

What is this? Observed burst counts with burst intensity overlay. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Observed count, Rate ratio.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Observed burst counts with burst intensity overlay.

Observed count: Bars show observed submissions in each burst window.
Rate ratio: Observed-to-expected count ratio per tested window.

Analysis Help

What is this? Burst Windows focuses on observed-vs-expected burst intensity. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Repeated high-rate-ratio windows with composition shifts at multiple window sizes rather than isolated spikes. Single-window rate-ratio peaks can be benign; stronger signals are contiguous runs of elevated rate ratios that recur at multiple window sizes (for example 5m and 30m both elevated). High observed counts with low q-values in sustained windows imply concentration beyond baseline expectation, especially when these bursts overlap with duplicate-name or swing anomalies. Suppressed or unusually flat burst activity can also be informative if baseline volume is high; a lack of natural variability may indicate synchronized intake behavior or batching. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate legitimate synchronized outreach or one-off reminder cascades. Momentary lows can indicate normal random fluctuation when expected baseline is already elevated. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate repeated concentration windows that deserve correlation with duplicate and swing signals. Extended lows can indicate suppressed variance that can indicate workflow smoothing or batching. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Single-window rate-ratio peaks can be benign; stronger signals are contiguous runs of elevated rate ratios that recur at multiple window sizes (for example 5m and 30m both elevated).
High observed counts with low q-values in sustained windows imply concentration beyond baseline expectation, especially when these bursts overlap with duplicate-name or swing anomalies.
Suppressed or unusually flat burst activity can also be informative if baseline volume is high; a lack of natural variability may indicate synchronized intake behavior or batching.

Common benign causes: Agenda release timing and outbound campaign alerts can generate short-lived legitimate bursts.

`bursts_significance_by_window`

Chart Help

What is this? Burst significance by window size. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Burst significance by window size.

Bar height: Number of significant windows for each tested window size.
X-axis: Window size in minutes.

`bursts_composition_shift`

Chart Help

What is this? Burst composition shift over time. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Observed count, Absolute pro-rate shift, Baseline pro rate, Delta pro rate, Low-power.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Burst composition shift over time.

Observed count: Observed submissions in each burst window.
Absolute pro-rate shift: Absolute deviation of burst-window pro rate from run baseline.
Baseline pro rate: Run-level baseline pro share for composition comparison.
Delta pro rate: Signed burst-window pro-rate shift from baseline.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

`bursts_null_distribution`

Chart Help

What is this? Burst null simulation output. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Burst null simulation output.

Bar height: Maximum simulated count observed in each null iteration.
X-axis: Simulation iteration index.

Pro/Con Swings

Chart Help

What is this? Pro-rate trend against baseline stability bands. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Pro rate, Wilson low / Wilson high, Baseline pro rate, Stable lower / stable upper, Flagged, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Pro-rate trend against baseline stability bands.

Volume: Bars show record volume in each time bucket.
Pro rate: Observed pro share in each bucket.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Baseline pro rate: Expected day/time pro share baseline.
Stable lower / stable upper: Expected range around baseline for normal fluctuation.
Flagged: Buckets flagged by swing detector for abnormal deviation.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Pro/Con Swings focuses on directional pro/con ratio movement relative to expected bands. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Sustained directional ratio changes across neighboring buckets and repeated dayparts, especially when contiguous same-direction runs lengthen. Brief pro-rate jumps with wide Wilson intervals typically indicate low-support noise; treat them as weak unless adjacent buckets move in the same direction with tighter intervals. Extended daytime streaks of positive or negative shifts (multiple contiguous buckets) can indicate directional mobilization, queueing effects, or operational gating; confirm with day/hour and time-of-day panels. Large off-hours directional blocks that reverse at wake-hour transitions may indicate temporally segmented participation behavior, including potential strategic timing by one side. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate small-sample randomness, especially in low-power buckets. Momentary lows can indicate brief balancing waves where opposite-side submissions cluster together. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate persistent directional mobilization or process-side skew in intake timing. Extended lows can indicate prolonged suppression of one side that may indicate queueing or campaign fatigue. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Brief pro-rate jumps with wide Wilson intervals typically indicate low-support noise; treat them as weak unless adjacent buckets move in the same direction with tighter intervals.
Extended daytime streaks of positive or negative shifts (multiple contiguous buckets) can indicate directional mobilization, queueing effects, or operational gating; confirm with day/hour and time-of-day panels.
Large off-hours directional blocks that reverse at wake-hour transitions may indicate temporally segmented participation behavior, including potential strategic timing by one side.

Common benign causes: Daypart participation mix and event-response waves can move ratios without manipulation.

`procon_swings_shift_heatmap`

Chart Help

What is this? Day/slot deviation heatmap. This is a matrix view where color encodes magnitude across paired axes such as date/hour or slot/day. Each cell is a compact summary of one intersection, so the chart is optimized for pattern shape over exact per-cell precision.

Why this matters: Heatmaps reveal spatially contiguous patterns that line charts can hide, especially repeated daypart behavior and slot-level drift. They are especially useful for finding regime-like blocks that persist across many adjacent cells.

How to interpret: Scan for contiguous blocks before focusing on single cells. Compare high-intensity and low-intensity regions with bucket support and related detector outputs. Then check whether color transitions occur at meaningful boundaries such as day changes, hearing windows, or slot shifts.

What to look for: Look for coherent blocks, repeated stripes, or abrupt regime boundaries that persist across adjacent rows/columns. Short isolated hot/cold cells are weaker evidence; long bands or rectangles are stronger. Legend components: Cell color, Slot outlier dots, Axes.

Momentary high/low: A single hot/cold cell can reflect transient activity or low support. Interpret isolated cells cautiously, especially if they do not repeat in neighboring slots. Momentary highs can map to one reminder wave; momentary lows can map to ordinary quiet periods.

Extended high/low: Extended hot/cold regions typically indicate sustained behavioral mode shifts. Persistence across multiple dates/slots is stronger evidence than one transition point. Extended hot regions may indicate durable mobilization or process bias; extended cold regions may indicate suppression or inactivity.

Legend guide: Day/slot deviation heatmap.

Cell color: Red cells are more pro-heavy than expected for that slot; blue cells are more con-heavy.
Slot outlier dots: Highlighted cells that exceed detector outlier thresholds.
Axes: X-axis is slot-of-day; Y-axis is calendar date.

`procon_swings_day_hour_heatmap`

Chart Help

What is this? Average pro-rate by weekday/hour. This is a matrix view where color encodes magnitude across paired axes such as date/hour or slot/day. Each cell is a compact summary of one intersection, so the chart is optimized for pattern shape over exact per-cell precision.

Why this matters: Heatmaps reveal spatially contiguous patterns that line charts can hide, especially repeated daypart behavior and slot-level drift. They are especially useful for finding regime-like blocks that persist across many adjacent cells.

How to interpret: Scan for contiguous blocks before focusing on single cells. Compare high-intensity and low-intensity regions with bucket support and related detector outputs. Then check whether color transitions occur at meaningful boundaries such as day changes, hearing windows, or slot shifts.

What to look for: Look for coherent blocks, repeated stripes, or abrupt regime boundaries that persist across adjacent rows/columns. Short isolated hot/cold cells are weaker evidence; long bands or rectangles are stronger. Legend components: Cell color, Axes.

Momentary high/low: A single hot/cold cell can reflect transient activity or low support. Interpret isolated cells cautiously, especially if they do not repeat in neighboring slots. Momentary highs can map to one reminder wave; momentary lows can map to ordinary quiet periods.

Extended high/low: Extended hot/cold regions typically indicate sustained behavioral mode shifts. Persistence across multiple dates/slots is stronger evidence than one transition point. Extended hot regions may indicate durable mobilization or process bias; extended cold regions may indicate suppression or inactivity.

Legend guide: Average pro-rate by weekday/hour.

Cell color: Darker cells indicate higher pro rate.
Axes: X-axis is hour of day; Y-axis is weekday.

`procon_swings_time_of_day_profile`

Chart Help

What is this? Pro-rate profile by slot-of-day. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Bar height, X-axis.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Pro-rate profile by slot-of-day.

Bar height: Pro share in that slot-of-day bucket.
X-axis: Slot start minute from midnight.

`procon_swings_direction_runs`

Chart Help

What is this? Contiguous pro/con directional runs over time. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, Line, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Contiguous pro/con directional runs over time.

Bar height: Number of contiguous buckets in each directional run.
Line: Mean absolute pro-rate shift magnitude across the run.
X-axis: Run start timestamp.

`procon_swings_null_distribution`

Chart Help

What is this? Null distribution for swing extremes. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Null distribution for swing extremes.

Bar height: Maximum absolute pro-rate delta per null iteration.
X-axis: Simulation iteration index.

Structural Changepoints

Chart Help

What is this? Volume/pro-rate timeline with structural break markers. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Pro rate, Wilson low / Wilson high, Flagged, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Volume/pro-rate timeline with structural break markers.

Volume: Bars show record volume in each time bucket.
Pro rate: Observed pro share over time.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Flagged: Detected changepoint locations.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Structural Changepoints focuses on structural breaks in level or composition. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Clusters of large-magnitude changes that align with other detector evidence windows. Look for clustered breakpoints across both volume and pro rate; multi-metric co-occurrence is usually more meaningful than a solitary break in one metric. Large absolute deltas with sustained post-break behavior (not immediate reversion) indicate structural transitions rather than transient spikes. Repeated changes at similar hours across days can reflect operational schedules; treat as lower risk unless change magnitudes are extreme and detector corroboration is strong. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate single regime boundaries caused by predictable hearing state transitions. Momentary lows can indicate noisy micro-fluctuations that do not persist across adjacent windows. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate multi-break episodes indicating stable before/after behavioral regimes. Extended lows can indicate a relatively stationary process with fewer systemic shifts. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Look for clustered breakpoints across both volume and pro rate; multi-metric co-occurrence is usually more meaningful than a solitary break in one metric.
Large absolute deltas with sustained post-break behavior (not immediate reversion) indicate structural transitions rather than transient spikes.
Repeated changes at similar hours across days can reflect operational schedules; treat as lower risk unless change magnitudes are extreme and detector corroboration is strong.

Common benign causes: Hearing open/close windows and coverage surges naturally create structural breaks.

`changepoints_magnitude`

Chart Help

What is this? Changepoint magnitude ranking. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Changepoint magnitude ranking.

Bar height: Absolute change magnitude at each detected break.
X-axis: Changepoint index/order.

`changepoints_hour_hist`

Chart Help

What is this? Changepoint timing histogram. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Changepoint timing histogram.

Bar height: Number of changepoints occurring in each hour-of-day bin.
X-axis: Hour of day (0-23).

Off-Hours Profile

Chart Help

What is this? Submission volume by hour-of-day. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Bar height, X-axis.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Submission volume by hour-of-day.

Bar height: Total submissions in each hourly bin.
X-axis: Hour of day (0-23).

Analysis Help

What is this? Off-Hours Profile focuses on overnight/off-hours participation and composition. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Consistent off-hours elevation in volume or composition beyond daytime baselines. A higher off-hours pro rate or off-hours volume share can be benign when support is low; prioritize differences that remain outside Wilson overlap and persist across days. Extended off-hours enrichment paired with normal daytime behavior can suggest time-targeted mobilization; compare with swing and periodicity detectors for repeated timing signatures. Unexpectedly low off-hours activity can also be anomalous in historically active datasets and may indicate ingestion gaps or narrowly timed campaign workflows. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate localized campaign pushes, timezone spillover, or delayed user activity. Momentary lows can indicate typical circadian troughs with naturally sparse submissions. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate systematic off-hours concentration that can indicate strategic timing. Extended lows can indicate consistently daytime-driven behavior and lower overnight engagement. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

A higher off-hours pro rate or off-hours volume share can be benign when support is low; prioritize differences that remain outside Wilson overlap and persist across days.
Extended off-hours enrichment paired with normal daytime behavior can suggest time-targeted mobilization; compare with swing and periodicity detectors for repeated timing signatures.
Unexpectedly low off-hours activity can also be anomalous in historically active datasets and may indicate ingestion gaps or narrowly timed campaign workflows.

Common benign causes: Statewide campaigns spanning time zones can shift participation into late-hour windows.

`off_hours_summary_compare`

Chart Help

What is this? Off-hours vs on-hours summary. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Off-hours vs on-hours summary.

Bar height: Off-hours pro-rate statistic for each summary row.
X-axis: Off-hours count/context grouping key.

`off_hours_day_hour_heatmap`

Chart Help

What is this? Off-hours composition by weekday/hour. This is a matrix view where color encodes magnitude across paired axes such as date/hour or slot/day. Each cell is a compact summary of one intersection, so the chart is optimized for pattern shape over exact per-cell precision.

Why this matters: Heatmaps reveal spatially contiguous patterns that line charts can hide, especially repeated daypart behavior and slot-level drift. They are especially useful for finding regime-like blocks that persist across many adjacent cells.

How to interpret: Scan for contiguous blocks before focusing on single cells. Compare high-intensity and low-intensity regions with bucket support and related detector outputs. Then check whether color transitions occur at meaningful boundaries such as day changes, hearing windows, or slot shifts.

What to look for: Look for coherent blocks, repeated stripes, or abrupt regime boundaries that persist across adjacent rows/columns. Short isolated hot/cold cells are weaker evidence; long bands or rectangles are stronger. Legend components: Cell color, Axes.

Momentary high/low: A single hot/cold cell can reflect transient activity or low support. Interpret isolated cells cautiously, especially if they do not repeat in neighboring slots. Momentary highs can map to one reminder wave; momentary lows can map to ordinary quiet periods.

Extended high/low: Extended hot/cold regions typically indicate sustained behavioral mode shifts. Persistence across multiple dates/slots is stronger evidence than one transition point. Extended hot regions may indicate durable mobilization or process bias; extended cold regions may indicate suppression or inactivity.

Legend guide: Off-hours composition by weekday/hour.

Cell color: Pro-rate for that weekday/hour cell (with low-power checks available in table data).
Axes: X-axis is hour of day; Y-axis is weekday.

Exact Duplicate Names

Chart Help

What is this? Exact-duplicate concentration over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Duplicate count, Duplicate count.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Exact-duplicate concentration over time.

Duplicate count: Exact duplicate occurrences per bucket.
Duplicate count: Bars show count of repeated canonical names in each bucket.

Analysis Help

What is this? Exact Duplicate Names focuses on exact repeated-name concentration. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? High duplicate concentration with frequent position switching for the same canonical name. Short bursts of repeated names in tiny windows may occur during legitimate group actions; concern rises when concentration repeats across multiple larger buckets. Names that appear repeatedly while switching pro/con positions are higher-priority review targets because they indicate inconsistent stance representation under one canonical identity. Persistent duplicate concentration during otherwise stable baseline periods can imply scripted submissions or queue replay effects rather than organic participation. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate household/shared-name collisions or small coordinated batches. Momentary lows can indicate normal diversity of distinct names in organic intake. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate repeat-name patterns likely to influence authenticity and weighting assumptions. Extended lows can indicate healthy name diversity with limited exact repetition pressure. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Short bursts of repeated names in tiny windows may occur during legitimate group actions; concern rises when concentration repeats across multiple larger buckets.
Names that appear repeatedly while switching pro/con positions are higher-priority review targets because they indicate inconsistent stance representation under one canonical identity.
Persistent duplicate concentration during otherwise stable baseline periods can imply scripted submissions or queue replay effects rather than organic participation.

Common benign causes: Common household names and family submissions may elevate duplicate counts.

`duplicates_exact_top_names`

Chart Help

What is this? Most repeated exact names. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Most repeated exact names.

Bar height: Total repeated occurrences for each display name.
X-axis: Top repeated display names.

`duplicates_exact_position_switch`

Chart Help

What is this? Repeated names with side switching. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Repeated names with side switching.

Bar height: Total records for repeated names appearing in multiple positions.
X-axis: Display names exhibiting pro/con switching.

Near-Duplicate Clusters

Chart Help

What is this? Near-duplicate cluster activity over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Cluster size, Records.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Near-duplicate cluster activity over time.

Cluster size: Bar height is cluster size at cluster first-seen time.
Records: Line shows total records represented by active clusters.

Analysis Help

What is this? Near-Duplicate Clusters focuses on near-duplicate cluster growth and similarity structure. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Large or fast-forming clusters with high edge similarity scores. Many small near-duplicate clusters are often benign typo/transliteration noise; stronger signals are rapid growth of large clusters in compressed time spans. High similarity edges among many distinct names during high-volume windows can indicate templated naming patterns or normalization collisions that warrant manual spot checks. Extended periods where cluster size and record counts rise together can point to coordinated intake streams, especially when aligned with burst and swing flags. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate short typo/transliteration bursts around outreach windows. Momentary lows can indicate periods where naming variation is naturally broader. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate templated or normalization-colliding naming behavior over sustained windows. Extended lows can indicate low cluster cohesion and reduced near-duplicate pressure. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Many small near-duplicate clusters are often benign typo/transliteration noise; stronger signals are rapid growth of large clusters in compressed time spans.
High similarity edges among many distinct names during high-volume windows can indicate templated naming patterns or normalization collisions that warrant manual spot checks.
Extended periods where cluster size and record counts rise together can point to coordinated intake streams, especially when aligned with burst and swing flags.

Common benign causes: Typos, OCR noise, and multilingual transliteration can inflate near-duplicate clusters.

`duplicates_near_cluster_size`

Chart Help

What is this? Cluster-size distribution. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Cluster-size distribution.

Bar height: Number of clusters observed at each cluster size.
X-axis: Cluster size (count of names).

`duplicates_near_time_concentration`

Chart Help

What is this? Near-duplicate cluster time concentration. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Bar height, X-axis.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Near-duplicate cluster time concentration.

Bar height: Peak time-bucket share of records for each near-duplicate cluster.
X-axis: Near-duplicate cluster identifier.

`duplicates_near_similarity`

Chart Help

What is this? Similarity levels among near-duplicate pairs. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Similarity levels among near-duplicate pairs.

Bar height: Similarity score for candidate name pairs.
X-axis: Left-hand name label from each pair sample.

Ordering / Sortedness

Chart Help

What is this? Ordering behavior across time buckets. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Records, Alphabetical indicator.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Ordering behavior across time buckets.

Records: Bar height is record count in each bucket.
Alphabetical indicator: Line values near 1 indicate alphabetical ordering for bucket windows.

Analysis Help

What is this? Ordering / Sortedness focuses on alphabetical/ordered submission behavior. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Bucket ranges with elevated alphabetical ratios and repeated minute-level ordering spikes. Single alphabetical spikes in small buckets can be accidental; repeated elevated alphabetical ratios across 15m-120m buckets suggest process-level ordering behavior. Sustained ordered streaks during high-volume windows are unusual for organic arrivals and may imply batch uploads, sorted lists, or deterministic queue processing. Low sortedness is expected for organic traffic, so abrupt transitions from unsorted to highly sorted and back are more informative than consistently modest ratios. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate small sorted snippets caused by chance or local administrative handling. Momentary lows can indicate expected unsorted arrivals from organic user behavior. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate batch-oriented or deterministic ordering processes across windows. Extended lows can indicate persistent organic ordering noise without process-level sorting artifacts. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Single alphabetical spikes in small buckets can be accidental; repeated elevated alphabetical ratios across 15m-120m buckets suggest process-level ordering behavior.
Sustained ordered streaks during high-volume windows are unusual for organic arrivals and may imply batch uploads, sorted lists, or deterministic queue processing.
Low sortedness is expected for organic traffic, so abrupt transitions from unsorted to highly sorted and back are more informative than consistently modest ratios.

Common benign causes: Batch exports or admin processing can produce temporary ordering artifacts.

`sortedness_bucket_summary`

Chart Help

What is this? Sortedness summary by bucket size. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Bar height, X-axis.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Sortedness summary by bucket size.

Bar height: Alphabetical ordering ratio for each bucket size.
X-axis: Bucket size in minutes.

`sortedness_kendall_tau_summary`

Chart Help

What is this? Kendall tau ordering strength by bucket size. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Kendall tau ordering strength by bucket size.

Bar height: Average absolute Kendall tau for each bucket size.
X-axis: Bucket size in minutes.

`sortedness_minute_spikes`

Chart Help

What is this? Minute-level ordering spikes. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Minute-level ordering spikes.

Bar height: Records seen in each minute-level ordering sample.
X-axis: Minute bucket timestamp.

Rare / Unique Names

Chart Help

What is this? Name uniqueness over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Unique ratio, Threshold unique ratio, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Name uniqueness over time.

Volume: Bars show record volume in each time bucket.
Unique ratio: Share of submissions with distinct canonical names per bucket.
Threshold unique ratio: Reference threshold used for unique-ratio anomaly signaling.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Rare / Unique Names focuses on novelty, uniqueness, and rarity concentration. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Sustained unique-ratio lifts with concurrent weirdness-score concentration. Short-lived unique-ratio increases during low volume can be misleading; investigate when unique-ratio elevation persists into higher-support windows. Concurrent rises in weirdness scores, singleton concentration, and rarity quantiles indicate novelty concentration beyond normal lexical drift. Extended rarity suppression (unusually low novelty) can also be noteworthy in broad public hearings and may suggest repeated template populations. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate brief novelty spikes from campaign expansion to new participants. Momentary lows can indicate common-name clustering or temporary shrinkage in participant diversity. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate sustained lexical novelty requiring cross-check against lookup coverage. Extended lows can indicate repeated-name dominance or limited participant turnover. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Short-lived unique-ratio increases during low volume can be misleading; investigate when unique-ratio elevation persists into higher-support windows.
Concurrent rises in weirdness scores, singleton concentration, and rarity quantiles indicate novelty concentration beyond normal lexical drift.
Extended rarity suppression (unusually low novelty) can also be noteworthy in broad public hearings and may suggest repeated template populations.

Common benign causes: Reference lookup gaps and nickname coverage gaps can overstate rarity.

`rare_names_weird_scores`

Chart Help

What is this? Highest weirdness-score names. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Highest weirdness-score names.

Bar height: Weirdness score of sampled names; higher indicates atypical string shape.
X-axis: Sample names sorted by weirdness.

`rare_names_singletons`

Chart Help

What is this? Singleton name composition over time. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Pro count, Con count.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Singleton name composition over time.

Pro count: Bars show pro-side count among singleton records.
Con count: Line shows con-side count among singleton records.

`rare_names_rarity_timeline`

Chart Help

What is this? Rarity-score timeline. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Rarity median, Rarity p95, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Rarity-score timeline.

Volume: Bars show record volume in each time bucket.
Rarity median: Median rarity score in each bucket.
Rarity p95: 95th percentile rarity score to show tail behavior.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Organization Field Anomalies

Chart Help

What is this? Blank organization-rate trend with position splits. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Volume, Blank org rate, Wilson low / Wilson high, Pro blank org rate, Con blank org rate, Low-power.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Blank organization-rate trend with position splits.

Volume: Bars show record volume in each time bucket.
Blank org rate: Overall blank/null organization share per bucket.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Pro blank org rate: Blank-org share among pro records.
Con blank org rate: Blank-org share among con records.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Organization Field Anomalies focuses on blank/null organization usage and split behavior. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Blank-rate surges that persist across higher-volume windows and one position side. Blank-organization spikes in low-support windows are weak evidence; prioritize wide windows where blank rate rises and Wilson bands remain narrow. Divergence between pro and con blank-org rates over sustained periods can indicate side-specific form behavior, campaign guidance, or data-entry heterogeneity. Sharp blank-rate reversals around specific times may indicate UX changes, batch imports, or conditional form paths and should be checked against operational logs. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate form UX friction or temporary omission guidance in outreach. Momentary lows can indicate short windows where organization prompts were more salient. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate systemic metadata sparsity that can bias affiliation interpretation. Extended lows can indicate more complete organization capture across participation streams. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Blank-organization spikes in low-support windows are weak evidence; prioritize wide windows where blank rate rises and Wilson bands remain narrow.
Divergence between pro and con blank-org rates over sustained periods can indicate side-specific form behavior, campaign guidance, or data-entry heterogeneity.
Sharp blank-rate reversals around specific times may indicate UX changes, batch imports, or conditional form paths and should be checked against operational logs.

Common benign causes: Form UX and campaign guidance often increase legitimate blank organization submissions.

`org_anomalies_position_rates`

Chart Help

What is this? Per-position blank-org rates by time bucket. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Blank org rate, Wilson low / Wilson high, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Per-position blank-org rates by time bucket.

Volume: Bars show record volume in each time bucket.
Blank org rate: Position-specific blank organization share.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

`org_anomalies_bursts`

Chart Help

What is this? Organization burst concentration. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Organization burst concentration.

Bar height: Burst count for organization-related minute windows.
X-axis: Minute bucket of organization burst sample.

`org_anomalies_top_orgs`

Chart Help

What is this? Most common organization values. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Most common organization values.

Bar height: Total records linked to each normalized organization value.
X-axis: Organization value labels.

Registered Voter Match

Chart Help

What is this? Registry match-rate trend over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Match rate, Wilson low / Wilson high, Pro match rate, Con match rate, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Registry match-rate trend over time.

Volume: Bars show record volume in each time bucket.
Match rate: Overall voter-registry name match rate per bucket.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Pro match rate: Match rate among pro-position records.
Con match rate: Match rate among con-position records.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Registered Voter Match focuses on name match coverage against voter registry reference. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Material and sustained match-rate departures with adequate per-bucket support. Transient match-rate drops in very small buckets are expected; treat as notable only when low-power flags are absent and the drop persists across neighboring windows. Sustained side-specific divergence (pro vs con) with adequate volume may indicate composition shifts, normalization mismatch, or targeted non-registered participation. Rapid oscillation between high and low match rates can suggest mixed data sources or ingestion inconsistencies; cross-check unmatched-name concentration for diagnostics. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate temporary concentration in highly matchable names. Momentary lows can indicate normal alias/normalization mismatch in sparse buckets. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate stable overlap with known-voter naming patterns. Extended lows can indicate persistent mismatch patterns requiring normalization and source review. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Transient match-rate drops in very small buckets are expected; treat as notable only when low-power flags are absent and the drop persists across neighboring windows.
Sustained side-specific divergence (pro vs con) with adequate volume may indicate composition shifts, normalization mismatch, or targeted non-registered participation.
Rapid oscillation between high and low match rates can suggest mixed data sources or ingestion inconsistencies; cross-check unmatched-name concentration for diagnostics.

Common benign causes: Name normalization variance and registration recency can reduce observed match rates.

`voter_registry_match_by_position`

Chart Help

What is this? Match rate by position grouping. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Match rate by position grouping.

Bar height: Registry match rate for each position label.
X-axis: Normalized position label.

`voter_registry_unmatched_names`

Chart Help

What is this? Most frequent unmatched names. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Most frequent unmatched names.

Bar height: Count of unmatched records for each canonical name.
X-axis: Canonical unmatched name values.

`voter_registry_position_buckets`

Chart Help

What is this? Position-specific match rates across time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Match rate, Wilson low / Wilson high, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Position-specific match rates across time.

Volume: Bars show record volume in each time bucket.
Match rate: Per-position registry match share by bucket.
Wilson low / Wilson high: Confidence band for the proportion metric; wider bands indicate higher uncertainty.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Periodicity

Chart Help

What is this? Clock-face minute concentration. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Clock-face minute concentration.

Bar height: Observed event count at each minute-of-hour bin.
X-axis: Minute of hour (0-59).

Analysis Help

What is this? Periodicity focuses on recurring timing structure across minute and lag spaces. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Narrow periodic peaks and elevated rolling overdispersion that recur over long spans and align across methods. Minor periodic peaks are normal in outreach-driven datasets; stronger signals appear when clock-face concentration, autocorrelation peaks, and spectrum peaks align. Narrow high-power peaks at specific periods (for example near exact campaign cadence intervals) can indicate automation or tightly scheduled reminders. Extended suppression of expected periodic structure in otherwise campaign-heavy contexts may imply missing intervals or preprocessing artifacts. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate single reminder cycles or one-time timed campaign sends. Momentary lows can indicate flat/noisy slots where periodic patterns are not dominant. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate repeated cadence signatures that may indicate automation or strict scheduling. Extended lows can indicate weak periodic structure consistent with more organic arrival timing. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Minor periodic peaks are normal in outreach-driven datasets; stronger signals appear when clock-face concentration, autocorrelation peaks, and spectrum peaks align.
Narrow high-power peaks at specific periods (for example near exact campaign cadence intervals) can indicate automation or tightly scheduled reminders.
Extended suppression of expected periodic structure in otherwise campaign-heavy contexts may imply missing intervals or preprocessing artifacts.

Common benign causes: Calendar reminders and regular campaign sends can produce expected periodic structure.

`periodicity_autocorr`

Chart Help

What is this? Autocorrelation by lag. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Autocorrelation by lag.

Bar height: Autocorrelation coefficient at each lag in minutes.
X-axis: Lag length in minutes.

`periodicity_spectrum`

Chart Help

What is this? Top spectral periods. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Top spectral periods.

Bar height: Spectral power for each candidate period.
X-axis: Detected period in minutes.

`periodicity_rolling_fano`

Chart Help

What is this? Rolling Fano overdispersion by window size. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Rolling Fano overdispersion by window size.

Bar height: Median rolling Fano factor for each window size.
X-axis: Rolling window size in minutes.

Multivariate Anomalies

Chart Help

What is this? Multivariate anomaly score and support over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Anomaly score, Anomaly score percentile, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Multivariate anomaly score and support over time.

Volume: Bars show record volume in each time bucket.
Anomaly score: Combined feature-space anomaly score for each bucket.
Anomaly score percentile: Percentile rank of anomaly score within this run.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Multivariate Anomalies focuses on joint anomaly score across multiple behavioral features. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? Consecutive high anomaly-score buckets supported by other detector flags. Single high anomaly buckets with low support can be model-noise; prioritize consecutive high-score windows with model eligibility and corroborating detector evidence. Joint excursions in volume, duplicate fraction, blank-org rate, and pro-rate shape are stronger than any one feature spike in isolation. Extended high-percentile stretches can indicate sustained behavioral mode changes; inspect top buckets and feature projection for which dimensions drive score elevation. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate single-bucket feature coincidence without sustained corroboration. Momentary lows can indicate brief reversion to feature-space baseline. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate multi-feature regime changes needing manual validation and context checks. Extended lows can indicate feature combinations staying near historically typical mixtures. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

Single high anomaly buckets with low support can be model-noise; prioritize consecutive high-score windows with model eligibility and corroborating detector evidence.
Joint excursions in volume, duplicate fraction, blank-org rate, and pro-rate shape are stronger than any one feature spike in isolation.
Extended high-percentile stretches can indicate sustained behavioral mode changes; inspect top buckets and feature projection for which dimensions drive score elevation.

Common benign causes: Correlated event shocks can move multiple features together without abuse.

`multivariate_top_buckets`

Chart Help

What is this? Top anomaly buckets (scatter). This scatter plot maps each bucket as a point in feature space, often with color and size as additional signals. It is a relationship view, showing joint behavior rather than a single metric over time.

Why this matters: Scatter views expose joint-feature structure, clusters, and outliers that are not visible in one-dimensional summaries. They help determine whether anomalies are isolated outliers or part of a broader feature-space regime.

How to interpret: Read axis meaning first, then evaluate whether outliers are isolated or part of a cluster. Use color/size encodings to understand confidence and support. Cross-reference extreme points with time-based charts to determine whether they are single events or repeated states.

What to look for: Look for detached point clouds, extreme tails, and dense anomaly clusters that align with flagged windows. A compact cluster far from baseline often carries more weight than one far-away point with low support. Legend components: Point position, Point color, Point size.

Momentary high/low: A single extreme point may be a one-off event or model artifact. Validate with timeline charts and table support counts. Momentary lows are usually returns toward baseline and are often benign unless paired with abrupt nearby outliers.

Extended high/low: Large persistent outlier clusters imply broad feature-space drift. Extended low-intensity clustering implies stable baseline behavior. Sustained dual-cluster structure can indicate mixed populations or alternating operational modes.

Legend guide: Top anomaly buckets (scatter).

Point position: X-axis is bucket volume; Y-axis is anomaly score.
Point color: Color reflects anomaly-score percentile rank.
Point size: Bubble size scales with bucket volume.

`multivariate_feature_projection`

Chart Help

What is this? Feature projection scatter. This scatter plot maps each bucket as a point in feature space, often with color and size as additional signals. It is a relationship view, showing joint behavior rather than a single metric over time.

Why this matters: Scatter views expose joint-feature structure, clusters, and outliers that are not visible in one-dimensional summaries. They help determine whether anomalies are isolated outliers or part of a broader feature-space regime.

How to interpret: Read axis meaning first, then evaluate whether outliers are isolated or part of a cluster. Use color/size encodings to understand confidence and support. Cross-reference extreme points with time-based charts to determine whether they are single events or repeated states.

What to look for: Look for detached point clouds, extreme tails, and dense anomaly clusters that align with flagged windows. A compact cluster far from baseline often carries more weight than one far-away point with low support. Legend components: Point position, Point color, Point size.

Momentary high/low: A single extreme point may be a one-off event or model artifact. Validate with timeline charts and table support counts. Momentary lows are usually returns toward baseline and are often benign unless paired with abrupt nearby outliers.

Extended high/low: Large persistent outlier clusters imply broad feature-space drift. Extended low-intensity clustering implies stable baseline behavior. Sustained dual-cluster structure can indicate mixed populations or alternating operational modes.

Legend guide: Feature projection scatter.

Point position: X-axis is log volume; Y-axis is pro rate.
Point color: Color shows anomaly score intensity.
Point size: Bubble size scales with bucket volume.

Composite Evidence Score

Chart Help

What is this? Composite risk score over time. This time-aligned view shows how the measured signal changes across chronological buckets. It is the primary lens for identifying sequence, duration, and coincidence with external events.

Why this matters: Time-series structure distinguishes transient spikes from sustained regime changes and helps align detector evidence by timestamp. Without duration context, it is easy to overreact to one-bucket noise and miss broad shifts.

How to interpret: Read left to right, compare volume with rate/score overlays, and pay attention to uncertainty bounds and low-power markers where available. When zoomed in, verify whether local extremes persist across neighboring buckets and remain visible at wider scales.

What to look for: Look for repeated peaks, troughs, trend breaks, and persistent drifts across adjacent windows. Patterns that recur at the same daypart across dates are usually stronger than one isolated wave. Legend components: Volume, Composite score, Low-power.

Momentary high/low: Short highs/lows can reflect event timing, random variance, or small-sample effects. Confirm with neighboring buckets before treating them as material anomalies. A momentary high near a known outreach time can be benign; a momentary low during expected peak periods may indicate data lag.

Extended high/low: Extended highs/lows are stronger indicators of behavioral shifts, especially when they persist across multiple bucket sizes and coincide with corroborating detector outputs. Extended highs may indicate sustained mobilization or systematic bias; extended lows may indicate prolonged inactivity or missing segments.

Legend guide: Composite risk score over time.

Volume: Bars show record volume in each time bucket.
Composite score: Aggregate score from multi-detector evidence overlap.
Low-power: Markers for buckets with insufficient support where rates can swing from noise.

Analysis Help

What is this? Composite Evidence Score focuses on cross-detector evidence overlap and prioritization. This section combines a hero chart, supporting charts, and tables to separate one-off noise from meaningful sustained behavior. Treat it as a detector notebook: start broad, then drill into specific windows with evidence context.

Why this data matters: This data matters because it changes how confident you should be in an anomaly narrative. Strong claims should come from persistent, well-supported patterns rather than isolated spikes. It also prevents both over-calling benign fluctuations and missing slow-burn anomalies that only emerge over longer runs.

How do I interpret this data? Read the hero chart first for the dominant temporal structure, then use detail charts to test whether the signal repeats across scales, dayparts, or subgroup splits. Use tables to verify exact values and support counts behind flagged windows. When uncertainty bands or low-power markers are present, discount single-window jumps unless they recur with stronger support.

What do I look for? High-score windows with overlapping detector flags and strong local support. High composite windows are most useful when evidence-count is high and signals come from independent detectors rather than one detector repeated across scales. Short isolated composite spikes can still be benign; extended elevated runs with overlapping burst/swing/changepoint/ML evidence are higher-priority review candidates. Very low composite scores during known high-activity periods can reveal under-sensitive detector settings or data-quality gaps and should trigger configuration review. Investigation priority should increase when multiple independent views tell the same story at the same time.

What could a momentary high/low mean? Momentary highs can indicate short-lived detector agreement around a local event. Momentary lows can indicate isolated detector activity without consensus evidence. Treat both cautiously when low-power flags are present. A practical rule: do not escalate on a single bucket unless a nearby table row and at least one companion chart support the same direction.

What could an extended high/low mean? Extended highs can indicate durable multi-detector agreement that should drive investigation priority. Extended lows can indicate broad detector disagreement suggesting mostly baseline behavior. Persistence across adjacent windows and corroborating detectors raises confidence that the shift is meaningful. Extended runs deserve timeline annotation and root-cause notes so later reviewers can separate operational context from suspicious behavior.

High composite windows are most useful when evidence-count is high and signals come from independent detectors rather than one detector repeated across scales.
Short isolated composite spikes can still be benign; extended elevated runs with overlapping burst/swing/changepoint/ML evidence are higher-priority review candidates.
Very low composite scores during known high-activity periods can reveal under-sensitive detector settings or data-quality gaps and should trigger configuration review.

Common benign causes: Major events can legitimately raise multiple detectors in the same period.

`composite_evidence_flags`

Chart Help

What is this? Evidence-flag composition. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Evidence-flag composition.

Bar height: Count of windows containing each detector flag.
X-axis: Detector evidence flag token.

`composite_high_priority`

Chart Help

What is this? Highest-priority composite windows. This categorical/ranked chart compares values across labels, groups, or parameter settings. It emphasizes composition and concentration instead of chronology.

Why this matters: Category comparisons show concentration, imbalance, and dominance patterns that can explain why timeline signals moved. They are often the fastest way to identify which subgroup is driving a detector outcome.

How to interpret: Sort by magnitude, compare head vs tail behavior, and relate category concentration to corresponding detector windows. Check both absolute values and relative spacing so you can distinguish true concentration from a uniformly low baseline.

What to look for: Look for heavy concentration in a few categories, abrupt drop-offs, or rare categories with disproportionately high values. A long flat tail with one or two dominant bars often indicates a targeted driver worth validating in tables. Legend components: Bar height, X-axis.

Momentary high/low: A single dominant category may come from one campaign event or local data artifact. Check whether the dominance repeats over time. Momentary category suppression can also happen when total volume is temporarily low.

Extended high/low: Persistent dominance/absence across many categories can indicate structural participation effects rather than random variation. Extended concentration deserves follow-up to determine whether it is policy-driven outreach, operational process, or suspicious patterning.

Legend guide: Highest-priority composite windows.

Bar height: Composite score for top-ranked windows.
X-axis: Window timestamp bucket.

How To Use This Report

Triage

Data Quality Warnings

Raw vs Dedup Metrics

Hearing Context Metadata

Deadline Ramp Metrics

Stance By Deadline Window

window_evidence_queue

record_evidence_queue

cluster_evidence_queue

Window Drilldown

Causative Rows

Exact Dup Names (Window)

Near-Dup Clusters (Window)

Runs + Weirdness Comparison

Baseline Profile

baseline_day_hour_volume

baseline_top_names

baseline_name_length_distribution

Burst Windows

bursts_significance_by_window

bursts_composition_shift

bursts_null_distribution

Pro/Con Swings

procon_swings_shift_heatmap

procon_swings_day_hour_heatmap

procon_swings_time_of_day_profile

procon_swings_direction_runs

procon_swings_null_distribution

Structural Changepoints

changepoints_magnitude

changepoints_hour_hist

Off-Hours Profile

off_hours_summary_compare

off_hours_day_hour_heatmap

Exact Duplicate Names

duplicates_exact_top_names

duplicates_exact_position_switch

Near-Duplicate Clusters

duplicates_near_cluster_size

duplicates_near_time_concentration

duplicates_near_similarity

Ordering / Sortedness

sortedness_bucket_summary

sortedness_kendall_tau_summary

sortedness_minute_spikes

Rare / Unique Names

rare_names_weird_scores

rare_names_singletons

rare_names_rarity_timeline

Organization Field Anomalies

org_anomalies_position_rates

org_anomalies_bursts

org_anomalies_top_orgs

Registered Voter Match

voter_registry_match_by_position

voter_registry_unmatched_names

voter_registry_position_buckets

Periodicity

periodicity_autocorr

periodicity_spectrum

periodicity_rolling_fano

Multivariate Anomalies

multivariate_top_buckets

multivariate_feature_projection

Composite Evidence Score

composite_evidence_flags

composite_high_priority

Name/Cluster Forensics

Top Repeated Names

Top Near-Dup Clusters

Methodology

Evidence Taxonomy

Profiling Coverage (Artifacts)

Interpretation Guardrails

`baseline_day_hour_volume`

`baseline_top_names`

`baseline_name_length_distribution`

`bursts_significance_by_window`

`bursts_composition_shift`

`bursts_null_distribution`

`procon_swings_shift_heatmap`

`procon_swings_day_hour_heatmap`

`procon_swings_time_of_day_profile`

`procon_swings_direction_runs`

`procon_swings_null_distribution`

`changepoints_magnitude`

`changepoints_hour_hist`

`off_hours_summary_compare`

`off_hours_day_hour_heatmap`

`duplicates_exact_top_names`

`duplicates_exact_position_switch`

`duplicates_near_cluster_size`

`duplicates_near_time_concentration`

`duplicates_near_similarity`

`sortedness_bucket_summary`

`sortedness_kendall_tau_summary`

`sortedness_minute_spikes`

`rare_names_weird_scores`

`rare_names_singletons`

`rare_names_rarity_timeline`

`org_anomalies_position_rates`

`org_anomalies_bursts`

`org_anomalies_top_orgs`

`voter_registry_match_by_position`

`voter_registry_unmatched_names`

`voter_registry_position_buckets`

`periodicity_autocorr`

`periodicity_spectrum`

`periodicity_rolling_fano`

`multivariate_top_buckets`

`multivariate_feature_projection`

`composite_evidence_flags`

`composite_high_priority`